Delphine
Pre-season buying decisions for footwear. A buyer sends one photo of a shoe that hasn't launched, and Delphine predicts weekly sell-through velocity per store cluster, sizes the order, and returns a buy/skip tier with a calibrated uncertainty band.
Footwear buyers commit to inventory 6–12 months before a shoe reaches a shelf, working from gut feel and last year's numbers. The expensive part isn't picking a bad shoe — it's ordering the right shoe in the wrong depth, for the wrong stores. Both errors cost margin in opposite directions: too shallow and the season sells out early at full price you never captured; too deep and the surplus clears at markdown.
Delphine predicts units per active store per week at (SKU, banner, store-cluster, quarter) grain, then converts that into a newsvendor-sized order quantity per tier. Every prediction carries a conformal-calibrated P20/P50/P80 band, and an explanation of why that band is wide.
Visual Peer Matching
CLIP embeds the product photo and finds the visually nearest silhouettes across a 2,718-image historical catalog. Their real velocity histories become model features — and the buyer sees the closest historical SKUs alongside the number, so the prediction is arguable rather than oracular.
Cross-Banner Cascade Modeling
The group runs a premium-to-takedown ladder: a silhouette peaks in one banner roughly two quarters before the takedown lands in the next. Delphine carries per-banner peer velocity, slope, and peak-quarters-ago as features — so it forecasts where a style sits on that ladder, not just whether it's selling.
Store-Cluster Granularity
Per-banner K-means clustering on foot traffic and sales mix, with K chosen by silhouette score. The same shoe moves 1.5–2× faster in a flagship than an outlet, so buy depth is allocated per cluster rather than spread evenly across the fleet.
Calibrated Uncertainty
MultiQuantile CatBoost produces P20/P50/P80, conformally calibrated per banner and cluster against a 60% coverage target. Buyers get a stated confidence range and a plain-language reason when it's wide — cold start, thin peer sample, or no strong visual match.
A third of our units sell at a markdown, at an average depth of ten percent off normal price — and that depth has widened every year since 2023. Buy depth is what moves it. At our own scale, taking five to ten percent off the markdown bill is worth seven figures a year, and we can measure it because we own both the model and the P&L it runs against.
The nearest commercial equivalent — a venture-backed apparel forecasting platform built on the same bet, product imagery plus external signals — raised US$25M and was acquired by an enterprise planning vendor in 2025. It is no longer sold on its own: the capability now ships inside a planning platform whose demand module alone lists at US$80,000–$250,000 a year, with implementation budgeted at 60–120% of the first-year subscription on top. That vendor's published benchmark for the category is savings of roughly 2% of revenue. One difference worth stating: they market social-sentiment and influencer signals as inputs. We tested those and removed them — against our own data they measured net negative.